A fine-grained 3D point cloud classification method and system based on metric learning

Abstract

The present invention discloses a fine-grained three-dimensional point cloud classification method based on metric learning. First, the fine-grained three-dimensional point cloud data set is divided into a training set, a support set and a test set, and the training set samples are expanded; a full convolutional network with only multiple convolutional blocks is constructed, and multiple convolutional layers are used to construct convolutional blocks to realize multi-level feature fusion, and a twin full convolutional network training method is used to complete the training of the full convolutional network; the trained full convolutional network is used as a feature extractor, and only a set of support sets need to be constructed to classify any test three-dimensional point cloud. The method of the present invention is an open set classification method, which does not add complex network modules and can classify any specified category of targets in the test phase.

Description

Technical Field

[0001] The present invention relates to the fields of computer graphics and computer vision technology, and in particular to a fine-grained three-dimensional point cloud classification method and system based on metric learning. Background Art

[0002] As the application of LiDAR in fields such as autonomous driving becomes more and more mature, new requirements have emerged for the processing of LiDAR 3D point cloud data. Most existing 3D point cloud classification algorithms are oriented towards general target classification, that is, distinguishing categories with different attributes. Fine-grained classification distinguishes subcategories under the same attribute category. Since the differences between subcategories are small, especially when objects are represented by point cloud data, the classification is more difficult.

[0003] Traditional point cloud feature extraction algorithms use feature algorithms such as clustering, density, and normal vectors to construct features. However, these feature algorithms only capture the underlying feature information of the point cloud and are difficult to support high-level semantic understanding tasks. Influenced by deep learning technology, many works have applied it to the research field of three-dimensional point cloud data processing. The three-dimensional point cloud of LiDAR has uneven distribution and disorder in spatial distribution, and the convolutional neural network that processes two-dimensional images requires structured data input. It is a direct method to project the three-dimensional point cloud onto different views or convert it into voxel expression and then process it using image methods, but this loses the spatial information of the three-dimensional point cloud and affects the feature extraction effect.

[0004] The PointNet deep network proposed by the researchers can learn global features directly from the original point cloud. Later, the PointNet++ network was developed to strengthen the extraction of local features of the point cloud and use fully connected layers to realize the classification of three-dimensional point clouds. Existing three-dimensional point cloud classification methods are capable of general target classification tasks, but they still have great challenges in the field of fine-grained classification. Patent application number CN202210562458.0 discloses a fine-grained three-dimensional point cloud classification method based on information enhancement. It captures fine-grained information through two levels: intra-layer context perception and inter-layer correlation enhancement. However, the densely connected multi-layer feature fusion makes the network large and complex, and the network is sensitive to classification categories, and the model needs to be retrained when the category changes. Summary of the invention

[0005] The main purpose of the present invention is to propose a fine-grained three-dimensional point cloud classification method based on metric learning, which does not add complex network modules and realizes the classification of different sub-category targets under the same attribute category.

[0006] The technical solution adopted by the present invention is:

[0007] A fine-grained three-dimensional point cloud classification method based on metric learning is provided, comprising the following steps:

[0008] S1. Fine-grained 3D point cloud dataset division: the fine-grained 3D point cloud dataset is divided into a training set, a support set and a test set using a balanced ratio method;

[0009] S2, training set sample expansion, flipping, rotating, and scaling the three-dimensional point cloud data samples in the training set to form three-dimensional point cloud data as new samples to add to the training set, expanding the number of training set samples;

[0010] S3, full convolutional network construction, including multiple convolutional blocks, each convolutional block consists of multiple convolutional layers, the output of each convolutional block is feature pooled and then channel fused, and the fused features are passed through the last convolutional layer to obtain the output features;

[0011] S4, twin full convolutional network training, the twin full convolutional network is composed of two full convolutional networks with shared parameters of the network convolution layer in step S3, randomly select two groups of three-dimensional point cloud data samples a and sample b from the training set, sample a and sample b are sampled and grouped respectively, and two groups of local point cloud regions Z are obtained a and Z b , the two sets of local point cloud regions Z a and Z b Input them into the twin full convolutional network respectively to obtain two sets of feature vectors f a and f b , according to the two sets of feature vectors f a and f b Calculate the contrast loss L C , the twin fully convolutional network is trained using gradient descent;

[0012] S5, constructing support feature vectors, performing sampling and grouping processing as described in step S4 on each 3D point cloud data sample in the support set to obtain a local point cloud region, and inputting the local point cloud region into the fully convolutional network trained in step S4 to obtain a set of support feature vectors corresponding to the 3D point cloud data samples in the support set;

[0013] S6. Classify the target fine-grained three-dimensional point cloud. Sampling and grouping the target fine-grained three-dimensional point cloud data samples to be classified in the test set are performed as described in step S4 to obtain the corresponding local point cloud area. The local point cloud area is input into the fully convolutional network trained in step S4 to obtain the feature vector of the target to be tested. The similarity between the feature vector of the target to be tested and each feature vector in the support set feature vector set is calculated one by one, and the support sample category corresponding to the support feature vector with the highest similarity is used as the predicted category of the fine-grained three-dimensional point cloud sample of the target to be tested.

[0014] Following the above technical solution, the balanced ratio method in step S1 is to divide the data set under each category in the fine-grained three-dimensional point cloud data set into a training set, a support set and a test set according to a ratio of 7:2:1.

[0015] Following the above technical solution, the flipping, rotation, and scaling transformation of the three-dimensional point cloud data samples of the training set in step S2 refers to randomly flipping the three-dimensional point cloud data horizontally and vertically, randomly rotating it from -45° to 45°, and randomly scaling it by 0.5 to 1.5 times.

[0016] Following the above technical solution, in step S2, the three-dimensional point cloud data samples of the training set are flipped, rotated, and scaled to form three-dimensional point cloud data as new samples, thereby expanding the number of training set samples to 5 to 20 times the original number.

[0017] According to the above technical solution, in step S4, sampling and grouping the three-dimensional point cloud data samples to obtain a local point cloud area refers to randomly sampling the three-dimensional point cloud data samples to obtain n sampling points, and taking the n sampling points as the following technical solution, in step S4, according to the two groups of feature vectors f a and f b Calculate the contrast loss L C The formula is as follows,

[0018]

[0019] Among them, y = 1 means that the two input samples a and sample b belong to the same category, y = 0 means that the two input samples a and sample b do not belong to the same category, d a，b are two sets of eigenvectors f a and f b The distance between them, α is the threshold set according to actual needs.

[0020] Following the above technical solution, the calculation of the similarity sim in step S6 is expressed as follows:

[0021]

[0022] Among them, f t is the eigenvector in the support set eigenvector set, f c is the feature vector of the sample to be tested in the test set, ||f t || 2 is the 2-norm.

[0023] Following the above technical solution, the two sets of feature vectors f a and f b The distance between a，b Take the cosine distance and calculate it as follows:

[0024]

[0025] Among them, f a and f b are the feature vectors of 3D point cloud data samples a and b, respectively, ||f a || 2 is the 2-norm.

[0026] The present invention also provides a fine-grained three-dimensional point cloud classification system based on metric learning, comprising:

[0027] The fine-grained 3D point cloud dataset partitioning module uses the balanced ratio method to divide the fine-grained 3D point cloud dataset into a training set, a support set, and a test set;

[0028] The training set sample expansion module is used to flip, rotate, and scale the three-dimensional point cloud data samples in the training set to form three-dimensional point cloud data, and add them to the training set as new samples to expand the number of training set samples;

[0029] The full convolutional network building module is used to build a full convolutional network, including multiple convolutional blocks. Each convolutional block consists of multiple convolutional layers. The output of each convolutional block is subjected to feature pooling and channel fusion. The fused features are passed through the last convolutional layer to obtain the output features.

[0030] The twin full convolutional network training module is used to construct a twin full convolutional network with two full convolutional networks that share parameters in the network convolutional layer. Two groups of 3D point cloud data samples a and b are randomly selected from the training set. Samples a and b are sampled and grouped respectively to obtain two groups of local point cloud regions Z. a and Z b , the two groups of local point cloud areas Z a and Z b Input them into the twin full convolutional network respectively to obtain two sets of feature vectors f a and f b , according to the two sets of feature vectors f a and f b Calculate the contrast loss L c , the twin fully convolutional network is trained using gradient descent;

[0031] A support feature vector construction module is used to sample and group each three-dimensional point cloud data sample in the support set to obtain a local point cloud area, and input the local point cloud area into a trained full convolutional network to obtain a support set feature vector set corresponding to the three-dimensional point cloud data sample of the support set;

[0032] The target fine-grained three-dimensional point cloud classification module is used to sample and group the target fine-grained three-dimensional point cloud data samples to be classified in the test set to obtain the corresponding local point cloud area, input the local point cloud area into the trained full convolutional network to obtain the feature vector of the target to be tested, calculate the similarity between the feature vector of the target to be tested and each feature vector in the support set feature vector set one by one, and take the support sample category corresponding to the support feature vector with the highest similarity as the predicted category of the fine-grained three-dimensional point cloud sample of the target to be tested.

[0033] The present invention also provides a computer storage medium, which stores a computer program that can be executed by a processor, and the computer program executes the fine-grained three-dimensional point cloud classification method based on metric learning described in the above technical solution.

[0034] The beneficial effects of the present invention are as follows: the present invention adopts a full convolutional network constructed by multiple convolutional blocks, which reduces the complexity of the model, and uses multiple convolutional layers to construct convolutional blocks to achieve multi-level feature fusion, which reduces the amount of calculation and achieves the fusion of underlying fine-grained features and high semantic features; using the trained full convolutional network as a feature extractor, only one set of support sets needs to be constructed to classify any test three-dimensional point cloud. The method of the present invention is an open set classification method, which does not add complex network modules and can classify any specified category of targets in the test phase. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0036] Figure 1 It is a flow chart of the fine-grained three-dimensional point cloud classification method based on metric learning of the present invention;

[0037] Figure 2 This is a diagram of the full convolutional network training;

[0038] Figure 3 This is a diagram of twin fully convolutional network training. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] The present invention utilizes a fully convolutional network constructed by multiple convolutional blocks to reduce the complexity of the model. At the same time, multiple convolutional layers are used to construct convolutional blocks to achieve multi-level feature fusion, which reduces the amount of calculation and realizes the fusion of underlying fine-grained features and high semantic features. By using the trained fully convolutional network as a feature extractor, only a set of support sets needs to be constructed to classify any test three-dimensional point cloud.

[0041] A fine-grained three-dimensional point cloud classification method based on metric learning in an embodiment of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0042] S1. Fine-grained 3D point cloud dataset division: The fine-grained 3D point cloud dataset is divided into training set, support set and test set using the balanced ratio method.

[0043] S2, training set sample expansion, flipping, rotating, and scaling the three-dimensional point cloud data samples in the training set to form three-dimensional point cloud data as new samples and add them to the training set to expand the number of training set samples.

[0044] S3, full convolutional network construction, including multiple convolutional blocks, each convolutional block consists of multiple convolutional layers, the output of each convolutional block is feature pooled and then channel fused, and the fused features are passed through the last convolutional layer to obtain the output features;

[0045] As the best embodiment, Figure 2 As shown, a network with 16 convolutional layers is used. Starting from the first convolutional layer, every 3 convolutional layers constitute a convolutional block. The first 15 convolutional layers constitute a total of 5 convolutional blocks. The output of each convolutional block is subjected to feature pooling and channel fusion. The fused features are passed through the last convolutional layer to obtain the output features. The feature pooling here can use maximum feature pooling or mean feature pooling. As a preferred embodiment, maximum feature pooling is used.

[0046] S4, twin full convolutional network training, such as Figure 3 As shown, the twin full convolutional network is composed of two full convolutional networks with shared parameters of the network convolutional layer in step S3. Two groups of three-dimensional point cloud data samples a and sample b are randomly selected from the training set, and sample a and sample b are sampled and grouped respectively to obtain two groups of local point cloud regions Z a and Z b , the two sets of local point cloud regions Z a and Z b Input them into the twin full convolutional network respectively to obtain two sets of feature vectors f a and f b , according to the two sets of feature vectors f a and f b Calculate the contrast loss L C, the twin fully convolutional network is trained using gradient descent.

[0047] S5. Support feature vector construction. Each three-dimensional point cloud data sample in the support set is sampled and grouped as described in step S4 to obtain a local point cloud area. The local point cloud area is input into the fully convolutional network trained in step S4 to obtain a set of support set feature vectors corresponding to the three-dimensional point cloud data samples in the support set.

[0048] S6. Classify the target fine-grained three-dimensional point cloud. Sampling and grouping the target fine-grained three-dimensional point cloud data samples to be classified in the test set are performed as described in step S4 to obtain the corresponding local point cloud area. The local point cloud area is input into the fully convolutional network trained in step S4 to obtain the feature vector of the target to be tested. The similarity between the feature vector of the target to be tested and each feature vector in the support set feature vector set is calculated one by one, and the support sample category corresponding to the support feature vector with the highest similarity is used as the predicted category of the fine-grained three-dimensional point cloud sample of the target to be tested.

[0049] The fully convolutional network constructed by the present invention has only 16 convolutional layers. The feature vectors of two training samples are extracted to calculate the contrast loss to train the network. There is no need to use a complex attention module to extract local features, which reduces the complexity of the model and improves the speed of point cloud classification while ensuring the classification accuracy.

[0050] As a preferred embodiment, the balanced ratio method in step S1 is to divide the data set under each category in the fine-grained three-dimensional point cloud data set into a training set, a support set and a test set according to a ratio of 7:2:1, so as to ensure that the number of categories in the training set, the support set and the test set is consistent. In specific implementation, the ratio of the training set to the other two sets is generally divided between 7:3 and 9:1, and the ratio of the support set to the test set is generally 2:1 or 1:1. It is an optimal solution to divide them into the training set, the support set and the test set according to a ratio of 7:2:1. This method will not cause uneven category distribution due to data division, and retains the category distribution information in the original overall data set.

[0051] As a preferred embodiment, the flipping, rotation, and scaling transformation of the training set 3D point cloud data samples in step S2 refers to performing horizontal and vertical random flipping, -45° to 45° random rotation, and 0.5 to 1.8 times random scaling on the 3D point cloud data. Therefore, as the optimal parameters, -15° to 15° random rotation and 0.8 to 1.2 times random scaling are performed.

[0052] As a preferred embodiment, in step S2, the three-dimensional point cloud data samples of the training set are flipped, rotated, and scaled to form the three-dimensional point cloud data as new samples. The number of training set samples is expanded to 5 to 20 times the original. When the expansion multiple is small, the trained network is in an underfitting state. Continuing to increase the expansion multiple significantly improves the performance. When the expansion multiple exceeds 10 times, the performance continues to improve slightly, but more training time is consumed. When the expansion multiple exceeds 20 times, the trained network is in an overfitting state, and the network performance even decreases. Therefore, as the optimal parameter, the number of training set samples is expanded to 10 times the original.

[0053] As a preferred embodiment, in step S4, sampling and grouping the three-dimensional point cloud data samples to obtain a local point cloud area refers to randomly sampling the three-dimensional point cloud data samples to obtain n sampling points, and taking the n sampling points as the above technical solution, in step S4, according to the two groups of feature vectors f a and f b Computing the contrast loss C The formula is as follows,

[0054]

[0055] Among them, y = 1 means that the two input samples a and sample b belong to the same category, y = 0 means that the two input samples a and sample b do not belong to the same category, d a，b are two sets of eigenvectors f a and f b The distance between them, α is the threshold set according to actual needs, and α is generally set to 0.5 by default to achieve a balance between the distinction effect and the training difficulty.

[0056] As a preferred embodiment, the calculation of the similarity sim in step S6 is expressed as follows:

[0057]

[0058] Among them, f t is the eigenvector in the support set eigenvector set, f c is the feature vector of the sample to be tested in the test set, ||f t || 2 is the 2-norm.

[0059] As a preferred embodiment, the two sets of feature vectors f a and f b The distance between a，b Take the cosine distance and calculate it as follows:

[0060]

[0061] Among them, f a and f bare the feature vectors of 3D point cloud data samples a and b, respectively, ||f a || 2 is the 2-norm.

[0062] The fine-grained three-dimensional point cloud classification system based on metric learning in the embodiment of the present invention is used to implement the above method embodiment. The system includes:

[0063] The fine-grained 3D point cloud dataset partitioning module uses the balanced ratio method to divide the fine-grained 3D point cloud dataset into a training set, a support set, and a test set;

[0064] The training set sample expansion module is used to flip, rotate, and scale the three-dimensional point cloud data samples in the training set to form three-dimensional point cloud data, and add them to the training set as new samples to expand the number of training set samples;

[0065] The full convolutional network building module is used to build a full convolutional network, including multiple convolutional blocks. Each convolutional block consists of multiple convolutional layers. The output of each convolutional block is subjected to feature pooling and channel fusion. The fused features are passed through the last convolutional layer to obtain the output features.

[0066] The twin full convolutional network training module is used to construct a twin full convolutional network with two full convolutional networks that share parameters in the network convolutional layer. Two groups of 3D point cloud data samples a and b are randomly selected from the training set. Samples a and b are sampled and grouped respectively to obtain two groups of local point cloud regions Z. a and Z b , the two groups of local point cloud areas Z a and Z b Input them into the twin full convolutional network respectively to obtain two sets of feature vectors f a and f b , according to the two sets of feature vectors f a and f b Calculate the contrast loss L C , the twin fully convolutional network is trained using gradient descent;

[0067] A support feature vector construction module is used to sample and group each three-dimensional point cloud data sample in the support set to obtain a local point cloud area, and input the local point cloud area into a trained full convolutional network to obtain a support set feature vector set corresponding to the three-dimensional point cloud data sample of the support set;

[0068] The target fine-grained three-dimensional point cloud classification module is used to sample and group the target fine-grained three-dimensional point cloud data samples to be classified in the test set to obtain the corresponding local point cloud area, input the local point cloud area into the trained full convolutional network to obtain the feature vector of the target to be tested, calculate the similarity between the feature vector of the target to be tested and each feature vector in the support set feature vector set one by one, and take the support sample category corresponding to the support feature vector with the highest similarity as the predicted category of the fine-grained three-dimensional point cloud sample of the target to be tested.

[0069] The present application also provides a non-transitory computer-readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a disk, an optical disk, a server, an App application store, etc., on which a computer program is stored, and the program implements the corresponding function when executed by the processor. The computer-readable storage medium of this embodiment is used to implement the fine-grained three-dimensional point cloud classification method based on metric learning of the method embodiment when executed by the processor.

[0070] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A fine-grained 3D point cloud classification method based on metric learning, characterized in that: The following steps are involved: S1. Fine-grained 3D point cloud dataset division: the fine-grained 3D point cloud dataset is divided into a training set, a support set and a test set using a balanced ratio method; S2, training set sample expansion, flipping, rotating, and scaling the three-dimensional point cloud data samples in the training set to form three-dimensional point cloud data as new samples to add to the training set, expanding the number of training set samples; S3, full convolutional network construction, including multiple convolutional blocks, each convolutional block consists of multiple convolutional layers, the output of each convolutional block is feature pooled and then channel fused, and the fused features are passed through the last convolutional layer to obtain the output features; S4, twin full convolutional network training, the twin full convolutional network is composed of two full convolutional networks with shared parameters of the network convolution layer in step S3, randomly select two groups of three-dimensional point cloud data samples a and sample b from the training set, sample a and sample b are sampled and grouped respectively, and two groups of local point cloud regions Z are obtained a and Z b , the two sets of local point cloud regions Z a and Z b Input them into the twin full convolutional network respectively to obtain two sets of feature vectors f a and f b , according to the two sets of feature vectors f a and f b Calculate the contrast loss L C , the twin fully convolutional network is trained using gradient descent; S5, constructing support feature vectors, performing sampling and grouping processing as described in step S4 on each 3D point cloud data sample in the support set to obtain a local point cloud region, and inputting the local point cloud region into the fully convolutional network trained in step S4 to obtain a set of support feature vectors corresponding to the 3D point cloud data samples in the support set; S6. Classify the target fine-grained three-dimensional point cloud. Sampling and grouping the target fine-grained three-dimensional point cloud data samples to be classified in the test set are performed as described in step S4 to obtain the corresponding local point cloud area. The local point cloud area is input into the fully convolutional network trained in step S4 to obtain the feature vector of the target to be tested. The similarity between the feature vector of the target to be tested and each feature vector in the support set feature vector set is calculated one by one, and the support sample category corresponding to the support feature vector with the highest similarity is used as the predicted category of the fine-grained three-dimensional point cloud sample of the target to be tested.

2. The fine-grained three-dimensional point cloud classification method based on metric learning according to claim 1 is characterized in that: The balanced ratio method in step S1 is to divide the data set of each category in the fine-grained three-dimensional point cloud data set into a training set, a support set and a test set in a ratio of 7:2:

1.

3. The fine-grained three-dimensional point cloud classification method based on metric learning according to claim 1 is characterized in that: In step S2, the flipping, rotating, and scaling transformation of the three-dimensional point cloud data samples of the training set refers to randomly flipping the three-dimensional point cloud data horizontally and vertically, randomly rotating it from -45° to 45°, and randomly scaling it by 0.5 to 1.5 times.

4. The fine-grained three-dimensional point cloud classification method based on metric learning according to claim 1 is characterized in that: In step S2, the three-dimensional point cloud data samples of the training set are flipped, rotated, and scaled to form three-dimensional point cloud data as new samples, thereby expanding the number of training set samples to 10 times the original number.

5. The fine-grained three-dimensional point cloud classification method based on metric learning according to claim 1 is characterized in that: In step S4, sampling and grouping the three-dimensional point cloud data samples to obtain local point cloud areas refers to randomly sampling the three-dimensional point cloud data samples to obtain n sampling points, and grouping the point cloud data using the K nearest neighbor method with the n sampling points as the center to obtain n local point cloud areas.

6. The fine-grained three-dimensional point cloud classification method based on metric learning according to claim 1 is characterized in that: In step S4, according to two sets of feature vectors f a and f b Calculate the contrast loss L C The formula is as follows, Among them, y = 1 means that the two input samples a and sample b belong to the same category, y = 0 means that the two input samples a and sample b do not belong to the same category, d a,b are two sets of eigenvectors f a and f b The distance between them, α is the threshold set according to actual needs.

7. The fine-grained three-dimensional point cloud classification method based on metric learning according to claim 1 is characterized in that: The calculation of the similarity sim in step S6 is shown as follows: Among them, f t is the eigenvector in the support set eigenvector set, f c is the feature vector of the sample to be tested in the test set, ||f t ||2 is the 2-norm.

8. The fine-grained three-dimensional point cloud classification method based on metric learning according to claim 5, characterized in that: The two sets of feature vectors f a and f b The distance between a,b Take the cosine distance and calculate it as follows: Among them, f a and f b are the feature vectors of 3D point cloud data samples a and b, respectively, ||f a ||2 is the 2-norm.

9. A fine-grained 3D point cloud classification system based on metric learning, characterized in that: include: The fine-grained 3D point cloud dataset partitioning module uses the balanced ratio method to divide the fine-grained 3D point cloud dataset into a training set, a support set, and a test set; The training set sample expansion module is used to flip, rotate, and scale the three-dimensional point cloud data samples in the training set to form three-dimensional point cloud data, and add them to the training set as new samples to expand the number of training set samples; The full convolutional network building module is used to build a full convolutional network, including multiple convolutional blocks. Each convolutional block consists of multiple convolutional layers. The output of each convolutional block is subjected to feature pooling and channel fusion. The fused features are passed through the last convolutional layer to obtain the output features. The twin full convolutional network training module is used to construct a twin full convolutional network with two full convolutional networks that share parameters in the network convolutional layer. Two groups of 3D point cloud data samples a and b are randomly selected from the training set. Samples a and b are sampled and grouped respectively to obtain two groups of local point cloud regions Z. a and Z b , the two sets of local point cloud regions Z a and Z b Input them into the twin full convolutional network respectively to obtain two sets of feature vectors f a and f b , according to the two sets of feature vectors f a and f b Calculate the contrast loss L C , the twin fully convolutional network is trained using gradient descent; A support feature vector construction module is used to perform the sampling and grouping processing described in step S4 on each three-dimensional point cloud data sample in the support set to obtain a local point cloud region, and input the local point cloud region into the fully convolutional network trained in step S4 to obtain a support set feature vector set corresponding to the three-dimensional point cloud data sample in the support set; The target fine-grained three-dimensional point cloud classification module is used to sample and group the target fine-grained three-dimensional point cloud data samples to be classified in the test set to obtain the corresponding local point cloud area, input the local point cloud area into the trained full convolutional network to obtain the feature vector of the target to be tested, calculate the similarity between the feature vector of the target to be tested and each feature vector in the support set feature vector set one by one, and take the support sample category corresponding to the support feature vector with the highest similarity as the predicted category of the fine-grained three-dimensional point cloud sample of the target to be tested.

10. A computer storage medium, characterized in that: A computer program executable by a processor is stored therein, and the computer program executes the fine-grained three-dimensional point cloud classification method based on metric learning as described in any one of claims 1 to 8.

Images